JPH02201607A - Adaptive control system using neural network - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] [Table of contents] Overview Industrial field of application Prior art (Figure 13) Means for solving the problems to be solved by the invention (Figure 1) Working examples (a) System Explanation (Figures 2 to 6) (b)
Explanation of neural network (Figures 7 to 11) (C) Explanation of specific examples (Figure 12) (d) Explanation of other embodiments Effects of the invention [Summary] Structure of a nonlinear system unknown controlled object: j , -ral1'.

Regarding adaptive control system 1, which uses a non-linear system to identify the network and make its output follow the output of a reference model, we have developed a non-linear system for the controlled object, and the function form of the non-linear part is unknown. Even in the case of and a control device which calculates an internal signal having the control output as a component using a neural machine work and generates control power from the calculation result and the IFI standard,
An identification device that calculates an identified value from the calculation result and a control output, and uses the control input as a teacher signal, and calculates the weight of the work from the error with the identified value. A learning device for calculating the control input and the control output, and a human camera for storing the control input and the control output and emitting the internal signal, Field] Akimi Suekan, the structure of a nonlinear system unknown@I object is a two-dimensional work, and the neural network is used to follow the output of the standard model. This invention relates to an adaptive control system-i-J- using

In cases where the varimeter to be controlled is unknown, or in cases where parameters fluctuate due to changes in the operating environment, an adaptive control method that performs control while identifying the varimeter to be controlled is useful.

As such an adaptive control force method 3, model reference adaptive control has been proposed in which the output of the control target 1 follows the output of the reference model.

However, the above-mentioned proposal is limited to linear control objects, and a satisfactory adaptive control method has not been established when the control object is a nonlinear system.

In the actual industrial field, the object to be controlled is rarely nonlinear (and it is difficult to imagine adaptive control technology that performs constant control on nonlinear systems).

[Conventional technology]

FIG. 13 is an explanatory diagram of the prior art.

Model-norm adaptive control, which is a powerful means of adaptive control,
As a method for extending the control object to a nonlinear system, as shown in FIG.
A method of estimating 1 to wn as parameters has been proposed.

For example, it is shown in the paper 1 Model Norm-Constrained Nonlinear System Model Normative Adaptive Control 1 on pages 55 to 58 of the 8th Adaptive Control Symposium Materials, 1988.

That is, a control device 2 and an identification device 3 are provided for the controlled object 1, a known nonlinear function system 1-n is provided in the control device 2 and the identification device 3, and the control human power U and identification In response to the identification error ε with the value u, the coefficients Wl-Wn of the system of the control device 2 and the identification device 3 are changed to perform identification and control so that the identification error ε becomes zero. .

[Problem to be solved by the invention]

In such conventional techniques, if the functional form of the nonlinear part itself is known, it is possible to identify and control it well.

However, when the functional form of the nonlinear part itself is unknown, identification and control cannot be performed, creating a problem in which it is difficult to apply the model reference adaptive control system to an actual controlled object.

Therefore, an object of the present invention is to provide an adaptive control system using a neural network that can identify and control a nonlinear system controlled object even when the functional form of the nonlinear part is unknown. shall be.

[Means for solving problems] FIG. 1 is a diagram showing the principle of the present invention.

As shown in FIG. 1, the present invention is an adaptive control system l that identifies and controls a controlled object 1 and causes the output of the controlled object 1 to follow a target value. Neural network 2 uses internal signals as components.
a control device 2 that calculates with a neural network 3a and generates control human power from the calculation results and a target value; an identification device 3 that calculates the internal signal with a neural network 3a and generates an identified value from the calculation results and the control output; a learning device 4 that uses the control input as a teacher signal and calculates the weights of the neural networks 2a and 3a from the error with the identified value; and a learning device 4 that stores the control input and the control output and generates the internal signal. It has an input/output memory 5.

[Effect]

In the present invention, since the neural network can approximate any continuous function, the control device 2 and the identification device 3
Neural networks 2a and 3a are provided for use in identifying nonlinear functions.

Then, the neural networks 2a and 3a are operated by the internal signals of the input/output memory 5, and the learning device 4 calculates and sets the weights of the neural networks 2a and 3a from the error, thereby making it possible to identify and control the nonlinear unknown control target. It is something to do.

〔Example〕

(a) System description FIG. 2 is a configuration diagram of an embodiment of the present invention.

In the figure, the same parts as those shown in FIGS. 1 and 13 are indicated by the same symbols, 2b and 3b are multipliers, 2c and 3C are adders, and 4a is a differencer. It is.

First, set up the problem as follows.

Assume that the controlled object 1 is a multi-input multi-output discrete-time nonlinear system expressed by the following equation.

y (t) -fo [y(t-1) 7. ++,
y(t-n) 7;u(t-d-1)'', -, u(
t-d-m) ']Mogo [y(t-1)T+-+y
(t-n)"; u(t-d-1)T+ -, u(t-
d--riu(t-d) force, output, and is a one-dimensional vertical vector.

d is a known dead time and is a constant. It is assumed that the orders n and m are known. re(・) and go(・) each have a nonlinear function as an element! Dimensional vertical vector, with nonlinear functions as elements! It is an xi-dimensional matrix, and the functional form of each element is unknown.

From the right side of equation (1), y(t-1), ..., y(t-d
), we get y(t)=fa[x(t-d)”l +g,+[x
(t-d)"] u(L-d) (2)x(t)"=
I-y(t-1)", -, y(t-n)"; u(t
-1)"s=,u(L-d-m)"l (3).

A vector x (t) whose components are the input and output of the controlled object is called an internal signal.

Here, d et (matrix) (gd[x(t-d) co]
Assuming that ≠0, from equation (2), the interconnection expression of the controlled object is u(t-d)=f [x(t-d)] 10g [x(t
-d)] y(t) (4)g [x(t-d
)1-g, [x(t-d)co-'(5)f [x(
L-d)] --gd [x(t-d)ko'ra
[x(t-d)'' (6) is obtained.

The control purpose is to calculate (5
), unknown interval #9 of equation (6). fc・), g(・), and convert the output V(t) of the controlled object into the output y,% of the exemplary model.
(1) shall be followed. However, the target (1!ym(y-+d)) which is d hours ahead in time is available.

Next, the identification device 3 will be explained.

The method for identifying the controlled object is as follows.

In order to identify the interconnection expression (4) of the controlled object, consider the following equation as an identification model.

u"(t-d)=f" [x(t-d)ko +g'
[x(t-d)] y(t) (7) Here,
In equation (7), u''(t-d), f''i), g''(・) are
u(t-d), f(・), g(・) in equation (4), respectively
is the estimated value.

In addition, identification error ε(t-d), identification squared error E(t-d)
is defined by the following equation.

ε(t-d) −u(t-d)-u”(t-d)
(8) p(t-d)-EεCt-d) ) 2/2
(9) FIG. 3 is an explanatory diagram of the four-layer neural network having the configuration shown in FIG. 2, and FIG. 4 is an explanatory diagram of the basic units constituting the neural network.

The identification device 3 has therein a neural network 3a that approximately realizes the nonlinear functions f(.) and g(.).

As shown in FIG. 4, the basic unit 30 of the neural network 3a is a sigmoid function σ(
) to perform the following operation.

neJ−Σ−jiXi〇 (10)yJ・−σ(
neLJ) (11) However, σ(netj=1/(1−+−exp(neLJ))
(12) Here, in order to facilitate the handling of the threshold value θ, 6°−〇, χ. -1, neJ-Σw;rx+ (13)y, -σ
(neJ) (14) and calculate the threshold θ
is treated as the same as weight.

The neural network 3a that approximately realizes the functional forms of r(.) and g(.) is constructed by layering basic units 30 in multiple stages, as shown in FIGS. 3(A) and 3(B).

Regarding the function f(・), as shown in Figure 3(A),
The output function of the human output layer is a linear function, and the two intermediate layers (
The output function of the hidden layer has a structure of a four-layer neural network 3aF in which both output functions are signoid functions.

Furthermore, the function g(·) has a structure 3aG in which similar four-layered networks are stacked in multiple stages.

Then, as shown in Fig. 2, (7) the input of the work is x(t-d), and the output is f spring [X
(t-d)) or g-jaku (x(t-d)).

Then, in order to execute equation (7), a multiplier 3b that calculates the product of matrix g (x) and vector V(t), and a vector f
An adder 3c for calculating the sum of (x) and g(x)·y is provided as shown in FIG. 2, and receives updated weights from the learning device 4.

Next, the learning device 4 is composed of a host computer,
It has a function of adjusting the weights of the neural network so as to minimize the identification squared error of equation (9) from the identification error ε(td) of the differentiator 4a.

It is f''1x(t-d)] of E(L-d), g
” [by backpropagation method using x(Ld) as the steepest method for 1.

FIG. 5 is a flowchart of weight adjustment processing.

First, in a four-layer neural network that approximately realizes the functional form of f(・), Output signal f,: i-th element ε of the first-order output signal vector of the output layer, i-th element of the co-identification error ε, +31.. Weight from the second intermediate layer to the output layer 12ゝ1.: Weight from the first hidden layer to the second hidden layer, N
), the weight Δ from the 8th layer to the 1st middle layer is +31
．． ．． ．． Culm 11. History new vertical 11, 21Ji, 11 (21, update amount Δ-C11,,
, 01, update amount η: Set as a learning constant.

The input/output relationship in each layer is expressed as si (using Bmoid function σ(・), rJ−Σ−”Jth”+ (17)hLZ) in equation (12),
= σ(Σ, +z+, h”'+ ) (18
)hflゝ −・−σ(Σ−tI)J, x+)
It can be written as (19).

13) Update of 1 is performed as follows.

1ηji(t+1)・w ”'; r (t)+6w
``'7i(t)-ηεJh(21, +21.' is updated as follows.

w ” i, (t + 1) = w ” 1 (t) + Δ, (2
1, , (L) However, -861w"□, (L) , (+1J, are updated as follows.

w"J8(t+1)=w"'j+(t)+ΔIf"'j
t(t) (31) However, the updating of the weights of the ral net 3aG is the same as in the case of r(·) except for equations (22), (23), and equations (28) to (30).

In (3) updating ji in g(・), gj
When lk is the (Ck) element of the matrix g5 of the output signal of the output layer, the equation (17) becomes gjk=Σw'', Hh''+(3B).

Equation (22) is: The above is the method for updating the weights of the four-layer neural network 3aF that approximately realizes the functional form of f(·).

Next, since it is a 4N new that approximately realizes the functional form of g(・) (where yk is the control output yOk-th element), 6w ''', +(t) = 778jy+i h '''
+ (23').

Equation (28) is: From equation (39), −Σ ε s Vkk4 ” sj (t)
(30').

The host computer (learning device) updates the weights at each time t, and sends the updated weights to the control device 2, the neural networks 2aF, 2aG, 3aF, and the identification device 3.
Transfer to 3aG.

Next, the control device 2 will be explained.

The control method for the controlled object 1 is performed as follows.

Since the yarn feeding to be controlled is expressed by equation (4), the identification device 3
The control input is given by the following equation using the estimation functions f''(.) and g&(.) realized by the neural networks 3aF and 3aG.

u(t)-f" [x(t)] +g attack[x(t)]
yn+(t+d) (40) This control input is x(t)
This is possible based on the content of

Therefore, the control device M2 has the same internal structure as the identification device 3 in FIGS. 3 and 4, receives and receives updated weights from the host computer 4, and stores input and output necessary for control as internal signals. The internal signal and the target value are received from the input/output memory 5, and the equation (40) is executed.

FIG. 6 is an explanatory diagram of the operation of the people count memory.

Comparing equations (7) and (40), we find that the unknown function f(・)
, g(·) is performed with a time delay of d and is a function of the internal signal X, so the input and output necessary for identification and control are stored as internal signals x(t) and x(t-d) However, the input/output memory 5 shown in FIG. 6 is required.

From equation (3), the input/output memory 5 stores past control outputs y(L-1) to y(t-n) in the storage area 5a of x(t),
Past control human forces u(t-1) to u(t-d-m) are stored and output as x(t).

Also, in the storage area 5b of x(t-d), from equation (3), the past control output y(t-d) is stored as x(t-d).
-4) ~y(t-d-n), control human power u(t-d-1
) to u(t-2d-+*), and store u(t-d) in the storage area 5C of u(t-d).

Therefore, the adaptive control system includes a turnout memory 5 for generating internal signals, and a turnout memory 5 (4) as shown in FIG.
A control device 2 that generates control human power u(t) according to equation (0), an identification device 3 that generates identified human power u''(t-d) using equation (7), and u (t-d) as a teacher signal. It consists of a learning device (host 1 - computer) 4 that updates the weights of a four-layer neural network.

The control device 2 and the identification device 3 are provided with neural networks 2aF, 2aG, 3aF, and 3aG.

A four-layer neural network with the siBmoid function as the output function, as shown in Figures 3 and 4, can realize any continuous function (for example, 1988 IEICE technical research report M B E 889). (Refer to the paper [On the reciprocal realization of continuous mapping using neural networks]).

In the present invention, this four-layer neural network is used to identify a nonlinear function with multiple inputs and multiple outputs.

At this time, subtracting equation (40) from equation (4), we get f[x(
t)] −f” [x(t)] 罎PX [x(
t)] y(t+d)g"[x(t)', l
ym(L+d)−〇 (41)r [x(t)] −
f” [x(t)−,14g [x(t)] [y(
t+d)-ym(t+d))+(g[x(t)]-
g” [x(t)] )ym(t+d)=O(42).

The neural network realizes the nonlinear function as a neighbor, and f
[x(t)] →f” [x(t)], g[
x(t)] →g”[x(t)] (t-+oO)
Then, from the assumption of det[g(・)IJ−0, y(t+d)→ym(t+d) (t→ω) (4
3), and the control objective is achieved.

Therefore, it is possible to identify and control a multi-input, multi-output controlled object 1 that includes any unknown nonlinear function.

(b) Explanation of Neural Network FIG. 7 is a block diagram of a neural network, showing the structure of the four-layer neural network shown in FIGS. 2 and 3.

In this example, the interface between the hierarchical structures of neural networks is realized in analog form.
The output of the basic unit 30 consists of analog signals.

In the figure, 31 is a weight output circuit provided for each basic unit 30, and a weight holding section (described later) of the basic unit 30.
32 is a weight signal line, which connects the output of the weight output circuit 31 to the weight holding circuit, and 33 is an initial signal output circuit provided according to the number of dimensions of the input button. , which outputs an initial signal that serves as an input pattern to the human-powered layer of the hierarchical network.

Reference numeral 34 denotes a synchronous control signal line, which transmits the same XJ1 control signal and weight from a main control circuit (to be described later) that controls data transfer to the weight output circuit 31, the initial signal output circuit 33, and the control circuit of the basic unit 30 ( This synchronous control signal line 34 is shown as a common line in the figure, but in detail, it is connected to the main control circuit for each circuit by an individual signal line. has been done.

35 is an analog lotus for transmitting the input signal to the input layer (the basic unit 30) and connecting the input layer and the middle layer, the middle layer and the middle layer, and the middle layer and the output layer;
Reference numeral 6 denotes a main control circuit, which outputs a synchronization control signal and weight according to instructions from the host computer 4.

Therefore, the four-layer neural network 1~ has a basic unit 30 for human input signals and a weight output circuit 31 in each layer from the input layer to the output layer, each layer is connected by an analog bus 35, and the main control circuit 36 The weights are set and synchronously controlled.

FIG. 8 is a configuration diagram of the basic unit.

The multiplication processing unit 302 in the figure is a multiplication type D/A converter 3
02a, the analog signal input from the basic unit 30 of the previous layer or the initial signal output circuit 33 (inputted via the input frame section 307), and the digital signal to be multiplied by that input. is multiplied by weight information (input via the weight holding unit 308), and the obtained multiplication result is output as an analog signal.

The accumulation processing unit 303 is composed of an analog adder 303a composed of an integrator and a sample hold circuit 303b that holds the addition result of the analog adder 303a.
The analog I-I adder 30.3a adds the output of the multiplier type D/A converter 302a and the previously obtained addition value held in the sample hold circuit 303h6 to generate a new addition value. , the sample and hold circuit 303b, which calculates the value, holds the added value calculated by the analog adder 303a, and feeds the held value to the analog adder 303a as the previous added value. The processing is executed in synchronization with the addition control signal output from the control circuit 309.

The dark value processing section 304 is composed of a nonlinear function generation circuit 304a that is an analog function generation circuit, and outputs a nonlinear signal such as a sigmoid function in response to an input. Multiplying type ■〕/Δ converter 3 for manually input analog signals
When the multiplication by 02 a and the accumulation of these multiplication results are completed, the output holding unit 305 performs arithmetic processing on the addition value X held in the sample hold circuit 303b to obtain the analog output value Y. is constituted by a sample and hold circuit, and holds the output value Y of the analog signal of the nonlinear function generation circuit 304a, which is output to the basic unit 30 in the subsequent layer.

Further, 306 is an output switch section, and a control circuit 309
By turning on for a certain period of time in response to the output control signal from
307 is an input switch unit that processes the final output held by the output holding unit 305 and outputs it onto the analog bus 35, and receives a manual control signal from the control circuit 309 and outputs it from the basic unit 30 in the previous layer. When the final output or the analog output from the initial signal output circuit 33 is sent, the input is accepted by turning ON, and this analog value is given to the multiplication type D/A converter 302a.

Reference numeral 308 denotes a weight holding unit, which is composed of a parallel outshift L-register, etc., and the weight signal sent from the main control circuit 36 is connected to the gate of the tri-state amplifier 308a (weight input control signal by the control circuit 309). is turned on), this weight signal is applied to the multiplication processing unit 3.
02 holds the necessary weights, and 309 is a control circuit that controls the processing functions of these basic units 1-30.

Multiplication processing section 302, accumulation processing section 303, and dark value processing section 3
As mentioned above, if the plurality of human power connected to the basic unit 30 is Yi, and the weight set for each connection is Wl, the multiplication processing unit 302 executes Y i −W i , the accumulation processing unit 303 calculates (-X(-θ)) will be calculated.

FIG. 9 is a block diagram of the main control circuit 36 configured in FIG. 7.

The main control circuit 36 in the figure includes an external bus interface circuit 36a, a microcode memory 36b, a program sequencer 36c, a control pattern memory 36d, and weight data 36e.

The external bus interface circuit 36a is connected to the host computer 4 via the main bus, and receives and receives operation instructions from the host computer 4. The microcode memory 36b is connected to the program sequencer 36c.
The program sequencer 36c stores microcode that defines the operation of the microcode memory 36.
control pattern memory 36 according to the microcode in b.
d and weight data 36e.

The control pattern memory 36d has its output signal line connected individually to each of the basic units 30 in the initial signal output circuit 33, input layer, intermediate layer, and output layer, and the control pattern memory 36d has its output signal line connected to each of the basic units 30 in the initial signal output circuit 33, input layer, intermediate layer, and output layer. For each group,
That is, for each set of initial signal output circuits 33, the set of human power layers, the set of a pair of intermediate layers, and the set of output layers, one circuit or basic unit 1 from each set is selected in a time-sharing manner. The weight data memory 36e, which is used to turn on and off the output signal line, is configured to set each basic unit 30 with a weight in synchronization with the time-sharing human input signal according to instructions from the program sequencer 36c. It outputs a weight (digital piggyback) to the weight output circuit 31.

FIG. 10 is an explanatory diagram of signal processing of a neural network.

When a request for conversion to an output pattern is given from the host computer 36 via the main bus, the main control circuit 36 cyclically sends an output control signal to the initial signal output circuit 33 in a time-series manner. The plurality of initial signal output circuits 33 are sequentially and cyclically selected in time series.

That is, the main control circuit 36 controls the program sequencer 3
6c, from the control pattern memory 36d,
First, in order to simultaneously give the basic unit 30 of the input layer -7-same town 1 system j wholesale signal) and to select the initial signal output circuits 33 one after another, the synchronous control signal line 34 is
Then, each initial signal output circuit 33 is turned on.

That is, first, in order to output the human pattern Y1 given to the initial signal output circuit 33 of "Ichiban" to the analog bus 35, the gate of the initial signal output circuit 33 among the nine synchronous control signal lines 34 is opened. Synchronous control signal line 34
(indicated by 34a-1 in the figure) is turned on, and the other synchronization control signal lines 34a are turned on.

Subsequently, the synchronous control signal line 3 opens the gate of the initial signal output circuit 33 in order to output the input pattern Y7 given to the next initial signal output circuit 33 to the analog bus 35.
4 (indicated by 34a-2 in the figure) is turned on, and the other synchronization control signal lines 34 are turned off.

Thereafter, the synchronous control signal line 34a is turned on and off in the same manner until the last input turn Yn of the initial signal output circuit 33 is output to the analog bus 35.

At the same time, in order to give a weight to each weight output circuit 31 of each basic unit 30 in the input layer, in synchronization with the ON operation of each synchronous control signal line 34a, each signal is transmitted via a synchronous control signal line 34b. Weight output times! At the same time, the output of the weight data memory 36e is set every 31 times.

In FIG. 10(A), the synchronization control signal of the synchronization control signal line 34a is represented by Yi output control signals (i-1 to n),
A process of cyclically selecting the initial signal output circuit 33 in time series is illustrated. Here, n is the number of initial signal output circuits 33.

The initial signal output circuit 33 selected in this way is connected to the analog lotus 35 (in the figure,
The input layer analog bus 35a) is processed so as to send out the analog signal Yi determined as an input layer.

Therefore, as shown in FIG. 10(A), the input layer 7-+-
The analog signals Yi are sent to the log bus 35a in an orderly manner corresponding to the number of initial signal output circuits 33, and
The first input pattern Yj, then the next input pattern Yi, and the next input pattern Yi, and so on, are repeatedly sent out one after another.

When the multiplication processing unit 302 of each basic unit 30 in the human power layer receives the analog signal Yi sent out, the weight holding unit 302, which is to be output from the main control circuit 36,
Using the weight Wi of 08, (Yi-W
i) will be executed.

Therefore, as shown in FIG. 10(B), the main control circuit 36 outputs the selected initial signal output circuit 33 via the weight output circuit 31 in synchronization with the selection process of the initial signal output circuit 33. The corresponding weight Wi is set in the weight holding section 308 of each basic unit 30 of the input layer.

This process of setting weights to the basic unit 30 can be realized according to either analog signal mode or digital signal mode.

2. Since the weight is specified for each connection, as mentioned above, it should be expressed more accurately as Wij (j is the basic unit number of the human power layer), or to simplify the explanation. That's why I chose Wi.

Here, the operation of the main unit I-3(+) will be explained according to the signal processing timing chart of the basic unit shown in FIG.

Note that here, the basic unit 30 in the input layer will be explained.

First, when the control circuit 309 receives a synchronization control signal given from the control pattern memory 36d of the main control circuit 36 via the synchronization control signal wA34a, the control circuit 309 receives a human control signal (c
) to turn on the manual switch unit 307 and at the same time open the gate of the 1-right state buffer 308a (d), the output switch unit 3
Turn on the output control signal (hl) that makes 06 conductive.

At this time, the main control circuit 36 sequentially turns on the above-mentioned synchronous control signal line 34a in synchronization with the clock (a), so the initial signal output circuit 36 synchronizes with this clock (a).
3a, 33b, and -33n are input to the analog bus 35 and the input switch section 307.
The signal is applied to the multiplication type D/A converter 302a via the multiplication type D/A converter 302a.

On the other hand, the main control circuit 36 similarly controls the weight data memory 3
Since the weight of 6e is given to the weight output circuit 31 via the synchronization control signal line 34b, this weight (digital data) Wi is stored in the weight holding unit 308 through the tri-state buffer 308a.

Also, at this time, the output control signal (hl) is set to the clock (a
) is turned on for the 111th term, so the basic unit is 30.
During this time, the analog gate of the sample and hold circuit of the output holding section 305 is in an open state, and the held analog value is output onto the intermediate layer analog bus 35 via the output switch section 306.

Now, when the weight Wl of the digital value is stored in the weight holding section 308, the multiplication control signal (e) is turned on, so that the multiplication type D/A converter 302a
The analog signal Y1 given through 7 and the weight Wl are multiplied by the weight Wl, and the multiplication result is output as an analog signal.

Subsequently, since the addition control signal CA) is turned on, the analog adder 303a consisting of an integrator operates, and the sample and hold cycle is completed! The analog value previously held in 303b (initially cleared and zero) is added to the multiplication result of the multiplication type D/A converter 302a, and the addition result is stored again in the sample and hold circuit 303b. be done.

With the above operation, one bus cycle is completed, and in synchronization with the next clock (a), the input switch section 307 gives the input button Y2 to the initial signal output circuit 33b, and the weight output circuit]1 gives this input button Y2. Since the weight W2 corresponding to the input cover turn Y2 is given, the input cover turn Y2 and the weight W
2 is performed, and the result of this multiplication is added to the hold value of the sample hold circuit 303b.

At this time, the output control signal (h2
) is turned on. From then on, this operation is repeated until the processing for the input turn Y7 of the initial signal output circuit 33n is completed.

Then, when the multiplication of the input cover turns Y, l and Wn is completed, the conversion control signal (g) is turned on, so that the value obtained by accumulating the multiplication results is the nonlinear function generation circuit 304a of the dark value processing section 304. , and the corresponding Y value is stored in the output holding unit 3.
It is held at 05. In other words, the dark value processing unit 304 performs the above-mentioned calculation process Y-1/(1+exp (-X10)), thereby obtaining the final output value Y, which is the final calculation output of the basic unit 30. output holding unit 30
5. When this value Y is determined, the accumulation processing unit 30
The accumulated value of 3 (the content held by the sample hold circuit 303b) is cleared by the input synchronization control signal in synchronization with the next selection cycle of the initial signal output circuit 33.

By performing the operations described above, each basic unit 30 obtains the final output value Y from the input cover turn Y and the weight W8.

From now on, the configuration of the embodiment shown in FIG.
I explain.

As explained in detail using Figure 11, the first jtJ of all-C
When the processing for the human cover 2' set in the 1 signal output circuit 33 is completed, the 1 control circuit 36 again gives a synchronization control signal to each basic unit 30 (d or n may be added as an identifier). Therefore, the input cover turn Y, which is newly sent to the initial signal output circuit 33, is
Similar operations are performed in accordance with the new gravity W applied from the main bus and the external HAS interface circuit 36a.

On the other hand, the final output value Y of the basic unit 30 of the input layer obtained in this way is held in the output holding section 3) 05, and is processed by analog The signal is sent via the bus 35 to the basic unit 30 of the intermediate layer located at the next stage in accordance with the time-division transmission format at -27.

That is, the main control circuit 36 controls each basic command of the human power layer.
1; Yo0 ,,,-, :7 □11 Control circuit 309 is supplied with C output control signal via t11 control line'(・ltl) (Fig. 11) 5-time j By sending out a cyclic U for copying purposes, the basic unit 30a ~
1,000 parts of output 1' of 30n (j is cyclically turned on in 2 time series. As a result, each basic unit 3
(The -1 relock signal of the final output value held in the output holding unit 305 of l a -3On is transmitted in a time-division manner to the multiplication processing unit 302 of each basic control unit 30a30n in the intermediate layer. It becomes -.

Each basic Yuyuno 'r' of the middle layer; Oa~30n is,
In the case of the previous i4i L, the same processing operation flame j7 is executed, the basic code of the middle layer obtained by this processing, ″'...1・30
The return output value Y of the basic unit 30 in the output layer can be found by executing the same time-division transmission process for the basic unit 30 in the intermediate layer using the final output value Y of b.
This will happen.

That is, the main control circuit 36 similarly controls each of the basic units J]1 to (1b,: individually connected synchronous control signal lines 34, 3.34b, of the intermediate layer and the output layer). each;(
6. Controls the basic units 30a to 30n.

FIG. 10(A) shows a timing chart of the output control signal to the basic unit 30 of the input layer in correspondence with the Y and output control signals to the initial signal output circuit 33, and also shows the timing chart of the output control signal to the basic unit 30 of the input layer. A timing chart of the final output value Y of the basic unit 30 of the human power layer that is sent to the analog lotus 35 (referred to as the middle layer analog lotus in the figure) which is sent to the analog lotus 35 (in the figure, referred to as the middle layer analog lotus) is shown below.

(C) Description of specific examples FIG. 12 is a detailed explanatory diagram of the present invention, and FIG. 12 (A
) shows an example of control of an inverted pendulum, and FIG. 12(B) shows an example of control of a link mechanism.

FIG. 12(A) shows the length 2! mounted on a smoothly moving trolley of mass M! , shows an inverted pendulum of quality ψm. Consider freely controlling the angle θ that the inverted pendulum makes with the vertical direction by applying force U to the cart. However, M, m, and E are unknown.

The equation of motion of an inverted pendulum is as follows.

y(L)・θ(1), dθ; dt −l:y(t+h) y<m)1/h
d20 /d f,”= [y(t+2h)−2y(
L+h)+y(L)l/h2 and difference approximation 1-2,
If we set k+n-tLnh, y(k+2)=2y(lu
I)-y(k)11(' Therefore, using the nonlinear functions f.(.) and Bo(.), it is expressed as a ζ-dimensional equation.

y(k)−fn [y(k−]), y(k−2)] +
go Iy(k-2), ] u(k-2) (48) Since this is in the form of the discrete-time nonlinear system (1) assumed by the present invention, applying the method of the present invention, A control target can be identified and controlled. Target value V of y(k),
If (k) is set to 0, the inverted pendulum can be balanced, and if ye(k) is set to a periodic signal, the inverted pendulum can be made to vibrate periodically.

In the two embodiments, one person can output one output, but multiple inputs and multiple outputs may be used.

FIG. 12(B) shows a link mechanism with two links.

Each link i is a homogeneous rod with mass m1 and length β. It is assumed that the control torque generated between the links is u1. θ is the i-th
Let the angle from the first link to the first link be a relative angle measured clockwise.

The equation of motion of the two-link system is as follows.

3(m+42mz)gffi + sinθ++3mz
g! 2sin(θ1θ2)+6u+
(49) 3mzgf25in(θ1+
θ2) +6oz (
50) However, L=2 [l111111"+razlz"+3Ta
z1M++lz CO2O3) ] (51)
Mz=2mz l 2” (52)y(t)−[θ
+(1), θz(t)] ”, u(t) = [u+
(t), uz(t)] When T is set and the same difference proximization as described above is performed, the following equation can be written using the nonlinear function vector f0(·) and the nonlinear function matrix go(·).

y(k)=fo [y(k-1)'', y(k-2)'']
+g6 [y(k-2)'] u(k-2) (53
) Since this also has the form of the discrete-time nonlinear system (1) assumed by the present invention, the method of the present invention can be applied.

Although an embodiment of the present invention for a two-link system has been shown here, the present invention can be similarly applied to a general n-link system.

(d) Description of other embodiments In the above embodiments, the neural network was constructed using hardware, but it can also be implemented using ProSensor software.

Further, although the neural network has been described as having four layers, it may have two or more layers.

Although the present invention has been described below with reference to two embodiments, various modifications can be made to the present invention in accordance with the gist of the present invention, and these are not excluded from the present invention.

〔Effect of the invention〕

As explained above, according to the present invention, since the control device and the identification device are configured using a neural network that can approximate any continuous function, adaptive control of a controlled object whose nonlinear part is unknown is possible. It has the effect of making it possible,
It greatly contributes to the construction of versatile control systems for various control objects.

[Brief explanation of the drawing]

Fig. 1 is a diagram of the principle of the present invention, Fig. 2 is a configuration diagram of an embodiment of the present invention, Fig. 3 is an explanatory diagram of the four-layer neural network of Fig. 2, and Fig. 4 is an illustration of the basic unit of Fig. 3. Explanatory diagram, Figure 5 is the second
Figure 6 is an explanatory diagram of the operation of the input/output memory in Figure 2, Figure 7 is a configuration diagram of the neural network in Figure 3, and Figure 8 is a configuration diagram of the basic unit in Figure 7. , Figure 9 is the seventh
Figure 1O is a diagram explaining the signal processing of the neural network in Figure 7. Figure 11 is a signal processing timing chart of the basic unit in Figure 8. Figure 12 is a detailed diagram of the present invention. Explanatory diagram: FIG. 13 is an explanatory diagram of the prior art. In the figure, 1-controlled object, 2-control device, 3-identification device, 2a, 3a-neural network, 4-learning device, input/output memory.

Claims (1)

[Claims] Identification and control are performed on the controlled object (1), and the controlled object (1) is identified and controlled.
In the adaptive control system that makes the output follow the target value in step 1), an internal signal whose components are past control inputs and control outputs is calculated by a neural network (2a), and a control input is generated from the calculation result and the target value. a control device (2); an identification device (3) that calculates the internal signal using a neural network (3a) and generates an identified value from the calculation result and the control output; and uses the control input as a teacher signal and generates the identified value. a learning device (4) that calculates the weight of the neural network (2a, 3a) from the error between the two; and an input/output memory (5) that stores the control input and the control output and generates the internal signal. An adaptive control system using a neural network, characterized by comprising: